The purpose of this study is to propose a novel prediction model for motor characteristics in the topology optimization (TO) of synchronous reluctance motors (SynRMs) using a Swin Transformer (ST) model. It was demonstrated that ST exhibits superior prediction accuracy compared to the convolutional neural network (CNNs) method for SynRMs. The attention mechanism that constitutes ST is employed to visualize the characteristic contribution region of the SynRMs.
The ST model was trained using datasets generated by TO. These datasets represent the material distributions in the SynRM rotors and their associated torque characteristics. The ST architecture uses a window-based, multi-head self-attention mechanism to capture global and local image features. The prediction accuracy of the average torque or peak-to-peak value of the torque was evaluated against the finite element method results, with CNNs serving as the baseline.
Compared with CNNs, the proposed method improves accuracy by up to 56.8% in terms of the mean square error of the average torque. Furthermore, the visualization method using the attention mechanism of ST effectively captured the material boundary features. The ST model has the potential to make accurate and interpretable predictions.
The proposed method constitutes a novel approach to the application of ST for the prediction of SynRMs. This approach addresses both predictive accuracy and explainability. The proposed method will be applied to TO and will extend the prediction targets to other characteristics of motors.
